Kia ora, Namaskaram đđž
What if AI could learn when to slow down and think again?
Researchers from IBM Research, University of Oxford, and Tulane University have built an AI system that does exactly that. Their architecture, called SOFAI (Slow and Fast AI), doesn't just think fast or slowâit develops metacognition: the ability to monitor its own thinking and decide when to switch gears.
In tests, SOFAI outperformed or matched both fast-only and slow-only approaches while demonstrating human-like skill learning, adaptability, and cognitive control.
What makes SOFAI different?
Inspired by Daniel Kahneman's Thinking, Fast and Slow, the teamâled by Francesca Rossi (IBM Research), Andrea Loreggia (University of Brescia), and Marianna B. Ganapini (Union College)âbuilt an AI with three cognitive layers:
đđž System 1 (S1): Fast, intuitive reasoning based on past experience
đđž System 2 (S2): Slow, deliberate reasoning that weighs evidence carefully
đđž Metacognition (MC): Monitors performance, decides which system to trust, and runs "what should I have done?" simulations to keep learning
Unlike earlier dual-process frameworks, SOFAI keeps these layers modular and explicit while making the switching logic transparent. SOFAI tracks its own confidence, computational cost, and context.
The Evidence
In complex environments, AI faces a trade-off between speed and accuracy. So the team tested SOFAI in a noisy grid-world with penalties for violating constraints.
The Goal: Find the shortest path that maximises reward while using as little computation time as possible (constraint).
Key results:
đđž Outperformed single-system approaches: Higher rewards and shorter routes with efficient time use.
đđž Learned through experience: Started cautiously with slow reasoning, then shifted to fast reasoning as it gained confidence.
đđž Adapted to complexity: Adjusted its approach based on S1 (fast thinking) reliability and environmental demands.
đđž Showed cognitive control: Under high risk aversion, it relied more on S2 (slow thinking)âslowing down and reflecting when stakes were high.
đđž SOFAI learns when to think fast and slow: Crucially, it learned when to think about its own thinking.
đ Reference
Bergamaschi Ganapini, M., Campbell, M., Fabiano, F. et al. (2025). Fast, slow, and metacognitive thinking in AI. npj Artificial Intelligence, 1, 27. https://doi.org/10.1038/s44387-025-00027-5
đđž Vishalâs Evidence-Based Prompt
Want to train your LLM to switch between fast and slow thinking like SOFAI?
Copy-paste this to the end of your prompt:
1ď¸âŁ Generate a fast mode answer:
Fast mode (System 1): Answer quickly based on intuitive pattern-matching from past experience.
2ď¸âŁ Evaluate switching criteria:
- Stakes if wrong? (Low/Medium/High)
- Confidence in fast answer? (0-100%)
- Multiple competing criteria to evaluate? (Yes/No)
Use Fast mode if: low stakes AND high confidence (at least 75%) AND single criterion
Switch to Slow mode if: medium/high stakes OR low confidence (less than 75%) OR multiple criteria
3ď¸âŁ If switching, regenerate using slow mode:
Slow mode (System 2): Deliberate step-by-step across multiple criteria (accuracy, diverse perspectives, unintended consequences).
4ď¸âŁ Explain your decision:
Which mode did you use and why?Designed with đ Vishal George
Founder & Chief Behavioural Scientist
